The data was used for validating CNN model performance for natural samples. The sample selection targeted on one sample per week from continuous operation between January to December 2021. Due to scarcity of some classes additional samples were selected from expected seasons. The selected samples were manually inspected: all classifications were assessed (confirmed or corrected) and all identifiable images that were left under the thresholds were labeled. The unidentifiable images that were left without an assigned class were considered as unclassified. More detailed explanation and example images can be found from the publication Kraft et al. 2022.
| The data set available here is published with article “Kraft et al. (2022). Towards operational phytoplankton recognition with automated high-throughput imaging, near real-time data processing, and convolutional neural networks. Front Mar. Sci. 9. Doi: 10.3389/fmars.2022.867695” and if used for further purposes, the article should be cited accordingly. The data set contains approximately 150 000 images belonging to 50 different classes (~57 000) + unclassifiable (~94 000) consisting mainly of phytoplankton. The images can be used to validate classifier model performance with data from natural samples. The images were collected with an Imaging FlowCytobot from a continuous deployment in 2021 at the Utö Atmospheric and Marine Research Station operated by Finnish Environment Institute and Finnish Meteorological Institute. The images were manually annotated by expert taxonomists. |